Skip to main content
POST
/
api
/
v1
/
evals
/
reward
Reward
curl --request POST \
  --url https://platform.composo.ai/api/v1/evals/reward \
  --header 'API-Key: <api-key>' \
  --header 'Content-Type: application/json' \
  --data '
{
  "messages": [
    {}
  ],
  "evaluation_criteria": "<string>",
  "evaluator_name": "<string>",
  "model_core": "align-20251111",
  "tags": {},
  "trace_id": "3c90c3cc-0d44-4b50-8888-8dd25736052a",
  "system": "<string>",
  "tools": [
    {}
  ],
  "evaluate_latest": true
}
'
{
  "explanation": "<string>",
  "score": 123,
  "sources": [
    {
      "id": "<string>",
      "type": "<string>"
    }
  ]
}

Authorizations

​
API-Key
string
header
required

Body

application/json

Request model for reward score evaluation of LLM responses against specified criteria.

​
messages
Messages · object[]
required

A list of chat messages in OpenAI or Anthropic messages format

Minimum array length: 2
​
evaluation_criteria
string | null

Criteria used for evaluation. Begins with one of the following: Reward, Penalize, Passes if, Fails ifNote, align-lightning-20251127 only supports Passes/Fails criteria. Mutually exclusive with evaluator_name.

Minimum string length: 1
​
evaluator_name
string | null

Name of an existing evaluator (group_id) to use for evaluation. Mutually exclusive with evaluation_criteria.

Required string length: 1 - 255
​
model_core
enum<string>
default:align-20251111

The model core for reward evaluation. Defaults to align-20251111 if not specified.

Available options:
align-20251111,
align-20250529,
align-20260109,
align-lightning-20251127,
align-lightning-20250731
​
tags
Tags · object

Optional key-value pairs to tag and categorize the request. All tag values must be strings.

​
trace_id
string<uuid> | null

Optional trace ID to associate this evaluation with an existing trace.

​
system
string | null

System message is separate for Anthropic-style LLM calls. Optional.

​
tools
Tools · object[] | null

List of tools available for the assistant to call in OpenAI or Anthropic format. Optional.

​
evaluate_latest
boolean
default:true

Whether to evaluate only the latest response or all responses. Only True is supported for align-lightning model cores.

Response

Successful Response

Response model for evaluation scores.

​
explanation
string
required

Explanation of the evaluation score

​
score
number | null

Evaluation score between 0 and 1. If null, the criteria was deemed not applicable.

​
sources
Source · object[] | null

Sources referenced in the evaluation, e.g. memories, annotations, or documents.